A Multi Fractal Model of Asset Returns

COWLES FOUNDATION FOR RESEARCH IN ECONOMICS
YALE UNIVERSITY
30 Hillhouse Avenue, New Haven CT 06520 8281
A Multifractal Model of Asset Returns

Benoit Mandelbrot∗
Department of Mathematics, Yale University and
IBM T. J. Watson Research Center
Adlai Fisher and Laurent Calvet†

Department of Economics, Yale University
Cowles Foundation Discussion Paper #1164
This Draft: September 15, 1997

First Draft: October 1996
THIS VERSION TEXT ONLY

Download figures at http://www.econ.yale.edu/∼fisher/papers.html
∗
10 Hillhouse Avenue, New Haven, CT 06520-8283. e-mail: fractal@watson.ibm.com
†
28 Hillhouse Avenue, New Haven, CT 06520-1972. e-mail: fisher@econ.yale.edu, lcalvet@minerva.cis.yale.edu
YALE UNIVERSITY
Abstract
This paper presents the multifractal model of asset returns (“MMAR”), based upon
the pioneering research into multifractal measures by Mandelbrot (1972, 1974). The
multifractal model incorporates two elements of Mandelbrot’s past research that are
now well-known in finance. First, the MMAR contains long-tails, as in Mandelbrot
(1963), which focused on Lévy-stable distributions. In contrast to Mandelbrot (1963),
this model does not necessarily imply infinite variance. Second, the model contains
long-dependence, the characteristic feature of fractional Brownian Motion (FBM), in-

troduced by Mandelbrot and van Ness (1968). In contrast to FBM, the multifractal
model displays long dependence in the absolute value of price increments, while price
increments themselves can be uncorrelated. As such, the MMAR is an alternative to
ARCH-type representations that have been the focus of empirical research on the distri-
bution of prices for the past fifteen years. The distinguishing feature of the multifractal
model is multiscaling of the return distribution’s moments under time-rescalings. We
define multiscaling, show how to generate processes with this property, and discuss
how these processes differ from the standard processes of continuous-time finance. The
multifractal model implies certain empirical regularities, which are investigated in a
companion paper.
Keywords: Multifractal Model of Asset Returns, Compound Stochastic Process, Sub-
ordinated Stochastic Process, Time Deformation, Trading Time, Scaling Laws, Multi-
scaling, Self-Similarity, Self-Affinity

YALE UNIVERSITY
1 Introduction
The probabilistic description of financial prices, pioneered by Bachelier (1900), initially focused on
independent and Gaussian distributed price changes. Financial economists have long recognized
two major discrepancies between the Bachelier model and actual financial data. First, financial
data commonly display temporal dependence in the alternation of periods of large price changes
with periods of smaller changes. Secondly, the tails of the histogram of observed data are typically
much fatter than predicted by the Gaussian distribution.
Inclusive of its several extensions, the ARCH/GARCH line of research, beginning with Engle
(1982) and Bollerslev (1986), has become the predominant mode of thinking about the statistical
representation of financial prices. In the original formulation of the GARCH(q, p) model, innova-
tions in returns are specified as
1/2
εt = ut ht ,
where ut are i.i.d., and
X
q X
p
ht = α0 + βj ht−j + αi ε2t−i .
j=1 i=1
Among the many extensions to this specification, Nelson (1991) generates asymmetric responses
to positive and negative shocks via more general functional dependence between ht and the past
values {(εt−n , ht−n )}. Other work adds an independent stochastic component to volatility itself.
Also, most of the recent literature weakens the i.i.d. assumption for {ut } to an assumption of
stationarity. Bollerslev, Engle and Nelson (1994) survey the literature thoroughly.
The common strand in GARCH-type representations is a conditional distribution of returns

that has a finite, time-varying second moment. This directly addresses volatility clustering in the
data, and mitigates the problem of fat tails.1
In our view, the most important topics in the recent GARCH literature include first, long
memory (Baillie, Bollerslev and Mikkelsen, 1996), and second, the relationship between statistical
representations at different time scales (Drost and Werker, 1996). These topics are central in
our multifractal model, and although the multifractal framework is substantially different than
1
Many studies find that additional weight in the tails is needed. This leads to the use of Student’s t-distributions,
Poisson jump components, or nonparametric representations of the conditional distribution.
1
YALE UNIVERSITY
GARCH, it is useful to briefly discuss the treatment of these problems in the familiar GARCH
setting.
Long-memory is, intuitively, the idea that the longest apparent cycle in a sample will be pro-
portional to the total number of observations. This idea has been formalized in several ways, most
commonly by a slower than exponential decay rate in the auto-correlation function. Hyperbolic
decay rates for the absolute value of asset returns were first reported by Taylor (1986) and are now
a well documented stylized fact of financial time-series.2
Figure 1 presents a visual display of the consequences of long-memory (or more accurately
the lack thereof). Figure 1a shows the first-differences of a simulated GARCH(1,1) process with
500 simulation periods. The parameters used in the model, {α1 , β1 } = {.05, .921}, come from
estimated parameter values in a study of US-UK exchange rates.3 This graph exhibits the type
of mild conditional non-stationarity characteristic of GARCH representations with low values of α

and high values of β.4 There are perceptible changes in volatility over relatively short (in terms of
number of lags) periods.
Figure 1b simulates many (100,000) periods. The previously noticeable short cycles disappear
between the peaks of the individual cycles. To the eye, this graph is indistinguishable from white
noise. This provides an example for a general rule: short-memory processes appear like white noise
from a distance.5
In contrast to Figure 1b, observed financial data contains noticeable fluctuations in the size
of price changes at all time scales. Figure 2 shows first-daily-differences in the logarithm of the
DM/US$ exchange rate from 1973 to the present. The most compelling aspect of this data is
the presence of temporal dependence of varying frequencies – a phenomenon which non-integrated

2
See Ding, Granger, and Engle (1993) and Dacorogna et al. (1993) for evidence, or Baillie (1996) for further
discussion.
3
See Mills (1993). pp 112-113. The study uses weekly data from 1980-1988, giving 470 observations.
4
The sum α + β is frequently referred to as the persistence in a GARCH(1,1) model. High α coefficients tend to
is a strong trade-off, because of the stability requirements of GARCH representations (e.g.

P i=1 αp +
P
produce sharper changes in volatility, while β contributes to more moderately evolving dependence in volatility. There
p q
j=1 βj < 1),
between these two types of coefficients. For this reason, the standard GARCH representation is unable to capture
large, immediate changes in volatility simultaneously with long cycles. Thus, in later literature, we see the application
of jump-processes or other independent stochastic shocks directly to the variance. This allows one to introduce sharp
changes in volatility without further restricting the permissible range of β. Nonetheless, even high values of β < 1
are subject to exponential decline in the ACF, which is at odds with the observed hyperbolic decline in some data.
Hence the recent interest in long memory processes.
5
This is a generalization of Donsker’s theorem to weakly dependent increments.
2
YALE UNIVERSITY
(weak-memory) GARCH-type representations do not capture.

Prior to the development of long-memory GARCH processes, researchers had limited alterna-
tives in modelling low frequency cycles. Attributing low frequency cycles to structural breaks,
effectively capping the sample span, is one option. However, since “structural breaks” potentially
pose great risk to investors, deliberate censoring of such events can lead to a serious underestimation
of market risks.
In theory, GARCH representations also offer the alternative of adding parameters of higher
orders to capture low frequency cycles. In practice, representing long-memory phenomena with a
non-integrated process leads to non-robust representations whose number of parameters grows with
the number of observations.
The recently developed FIGARCH process of Baillie, Bollerslev and Mikkelsen (1996), achieves
long memory parsimoniously. Like GARCH, FIGARCH has an infinite order ARCH representation
in squared returns. The model can be viewed as a set of infinite-dimensional restrictions upon its
ARCH parameters. Mathematically, the restrictions are transmitted by the fractional differencing
operator. In contrast to the FBM and ARFIMA, fractional differencing affects squared errors
rather than the error term itself. Hence, the martingale property of prices can be maintained
simultaneously with long memory in the absolute value of returns.
To visually reinforce the consequences of long memory, Figure 3a shows 500 simulated first
differences of a FIGARCH(1, d, 0) process. The parameter values are taken from Baillie, Bollerslev,
and Mikkelsen’s study of DM/USD exchange rates. When we repeat the simulation over 100,000
periods in Figure 3b, we see a variety of long-run variation in volatility. Unlike GARCH, there is
no convergence to Brownian behavior over very long sampling intervals.
A currently unexplored area in this new branch of the literature is the relationship between
FIGARCH representations at different time scales. Drost and Nijman (1993) and Drost and Werker
(1996) have studied this problem for GARCH. For a given class of discrete processes, they consider
temporal aggregation of log returns, obtaining processes defined on coarser time scales. If the
aggregated processes all belong to the same class as the original processes, they refer to the class
as closed under temporal aggregation. Another possible term for this property is scale-consistency,
since it implies an equivalence between representations of the model at different time scales. In
empirical work, lack of scale-consistency implies that the researcher adds an additional restriction
3
YALE UNIVERSITY
to the model when choosing the time-scale of the data.

The original GARCH assumptions, which Drost and Nijman call strong GARCH, are not scale-
consistent. The weak-GARCH class, which is scale-consistent, assumes only that the parameters
{α, β} are the best linear predictors in terms of lagged values of ht and ε2t , and that {ut } are
stationary. The exact distribution of the {ut } will, in general, be quite complicated to calculate
after aggregation, and correlation between the {ut } is unmodelled. For a continuous-time diffusion
with a weak GARCH(1, 1) representation, there is a unique correspondence between values of
{α, β} at different time scales. Drost and Werker suggest that estimates which fail to adhere to
this correspondence should be considered jump-diffusions.
The recent interest in long-memory and in scale-consistency within the GARCH literature
foreshadows two fundamental concepts in the Multifractal Model of Asset Returns (“MMAR”).
Multifractal processes will be defined by a restriction on the behavior in their moments as the
time-scale of observation changes. Like Drost and Werker, we will argue that information contained
in the data at different time scales can identify a model. Reliance upon a single time scale leads to
inefficiency, or worse, forecasts that vary with the time-scale of the chosen data.
Section 2 discusses previous financial models with scaling properties. Section 3 briefly introduces
the mathematics of multifractal measures and processes. Section 4 applies the idea of multiscaling
to financial time series, and presents the Multifractal Model of Asset Returns. Section 5 concludes.
This paper is the first in a three paper series that introduces the concept of multifractality
to economics. Like many ideas, multifractality can be understood at several different levels of
abstraction, and corresponding mathematical elegance. In this paper, we focus on a very concrete
aspect of multifractality - a scaling property in moments of the process or measure. In fact, we
define multifractals via this property in Section 3. This has two advantages. First, this definition
leads directly to an empirical test. Second, this definition avoids several mathematical technicalities
that are not necessary in an expository paper. Thus, we give a simple definition of multifractality,
present some examples, extend multifractality from measures to processes (which has not yet been
addressed in the mathematics literature), develop a model of financial price changes (the MMAR),
and discuss interesting properties of the MMAR from an economic perspective.
The second paper in the series, Calvet, Fisher, and Mandelbrot (1997), approaches the theory of
multifractals from an entirely different perspective. It focuses on the local properties of multifractal
4
YALE UNIVERSITY
processes, which substantially differ from the standard assumptions in continuous-time finance. In
particular, most diffusions are characterized by increments that grow locally at the rate (4t)1/2
throughout their sample paths. The exceptions, such as fractional Brownian motion, have local
growth rates of order (4t)H , where H is invariant over time. Multifractals, on the other hand,
have a multiplicity of local growth rates for increments, which leads to a quite elegant mathematical
theory. We explore this aspect of multifractals in the second paper, and develop concepts such as the
multifractal spectrum, which characterizes the distribution of local growth rates in a multifractal
process.
The third paper in the series, Fisher, Calvet and Mandelbrot (1997), is an empirical study,
which tests for multifractality in Deutschemark - US Dollar data. We find strong evidence of a
multifractal scaling law, and estimate the multifractal spectrum of the Deutschemark - US Dol-
lar time series. This allows us to recover MMAR components, and simulate a multifractal data
generating process. Further, we show that alternatives such as GARCH and FIGARCH are distin-
guishable from multifractals under simulation, and that the behavior of the exchange-rate data is
more consistent with the MMAR hypothesis.
2 Roots of the MMAR

2.1 Three Earlier Themes
While the MMAR is entirely new to economics, it combines three elements of Mandelbrot’s pre-
vious research that are now well-known. First, the MMAR incorporates long tails, although in
a substantially different form than Mandelbrot (1963), which focused on L-stable distributions.6
At the time of its publication, some researchers reacted against the L-stable model, principally on
the grounds that it implied an infinite variance.7 Most economists now agree that there is no a
priori justification for rejecting infinite second moments. Indeed, the L-stable model has recently
been applied and tested on foreign exchange and stock prices. Contributions to this literature in-
clude Koedijk and Kool (1992), Belkacem, Lévy-Véhel and Walter (1995), Phillips, McFarland and
McMahon (1996). While the MMAR accounts for long tails in financial data, it does not necessarily
imply an infinite variance of returns over discrete sampling intervals.

6
These distributions have alternately been called Lévy-stable, stable, Pareto-Lévy, and stable-Paretian.
7
See Cootner (1964).
5
YALE UNIVERSITY
Second, the multifractal model contains long-dependence, the characteristic feature of fractional
Brownian motion (FBM), which was formally introduced by Mandelbrot and van Ness (1968). FBM
has received wide applications in the natural sciences, particularly in hydrology. Its use in economics
was advanced by the work of Granger and Joyeux (1980) and Hosking (1981), and led to popular
models such as ARFIMA. Baillie (1996) provides a good review of this literature. The present
model builds on the FBM by obtaining long-memory in the absolute value of returns8, but allows
the possibility that returns themselves are white.
The remaining essential component of the multifractal market model is the concept of trading
time, introduced by Mandelbrot and Taylor (1967). The salient feature of trading time is explicit
modelling of the relationship between unobserved natural time-scale of the returns process, and
clock time, which is what we observe. Trading time models have been extensively used in the liter-
ature, including Clark (1973), Dacorogna et al. (1993), Müller et al. (1995), Ghysels, Gouriéroux
and Jasiak (1995, 1996).
All three of these components are contained in the MMAR. This model accounts for the most
significant empirical regularities of financial time-series, which are long tails relative to the Gaussian
and long memory in the absolute value of returns. At the same time, the model incorporates scale-
consistency, in the sense that a well-defined scaling rule relates returns over different sampling
intervals.
2.2 Self-Affine Processes
Returning to the question of how time should be discretized, Mandelbrot (1963, 1967), followed by
Fama (1963), suggested that the shape of the distribution of returns should be the same when the
time scale is changed. The property of invariance (up to a scale parameter) under aggregation of
independent elements has been studied by Paul Lévy (1925, 1937), and is called L-stability. The
class of L-stable random variables contains Gaussians and a continuum of distributions that have
scaling (Paretian) tails: P(X > x) ∼ Cx−α , with 0 < α < 2. These random variables can be used
to generate continuous-time random motions. Thus the L-stable processes, used in Mandelbrot
(1963), have stationary and independent stable increments. They include the Brownian Motion as
8
Taqqu (1975) establishes that a FBM BH (t) has long memory in the absolute value of its increments when
H > 1/2.
6
YALE UNIVERSITY
the only member with continuous sample paths. We can in turn generalize these processes to allow
for dependence in the increments.
Definition 1 Granted X(0) = 0, a random process {X(t)} that satisfies:

d
{X(ct1 ), ..., X (ctk )} = cH X(t1 ), ..., cH X(tk ) .
for some H > 0 and all c, k, t1 , ..., tk ≥ 0, is called self-affine.
We call H the self-affinity index, or scaling exponent, of X(t).
Authors such as Samorodnitsky and Taqqu (1994), refer to a process satisfying the above
definition as self-similar. We use the term self-similarity in a stricter sense, reserving it for geometric
objects which are invariant under isotropic contraction. Self-affinity is a more general term, which
allows for different rescalings along the directions of an orthonormal basis. Self-similarity applies
when the re-scaling operators are the same in each direction, so that the object is not only invariant
under dilations, but also rotations.
The distinction between self-similarity and self-affinity was drawn in Mandelbrot (1977). The
class of strictly (isotropically) self-similar stochastic processes is degenerate. However, among the
broader class of geometric objects and measures, the distinction between self-similarity and self-
affinity is essential.
Two main types of self-affine processes have been used in finance. The L-stable motions dis-
cussed above assume independent and stable increments. They contrast with Fractional Brownian
Motions (FBM), a class of self-affine processes BH (t) with continuous sample paths and Gaussian
increments. The exponent H, called self-affinity exponent, satisfies9 0 < H < 1. The FBM is a
Brownian Motion in the special case H = 1/2. For other values of H, the FBM has dependent incre-
ments. Autocorrelation is negative (anti-persistence) when 0 < H < 1/2, and positive (persistence)
when 1/2 < H < 1. Persistent FBM have long memory.
3 Multifractal Measures and Processes
L-stable processes miss one the main features of financial markets – the alternation of periods of
large price changes with periods of smaller changes.10 In contrast, the FBM is useful for modelling
9
Authors sometimes consider the degenerate case H = 1. The process is then of the form tZ, where Z is a normal
random variable (Samorodnitsky and Taqqu, 1994.)
10
This phenomenon is typically described as the fluctuation of “volatility” over time. In situations where variance
may be infinite, “volatility” should be taken to mean fluctuations in the expected absolute value of price changes.
7
YALE UNIVERSITY
the tendency of price changes to be followed by changes in the same (or opposite) direction. The
fractional Brownian Motions however, capture neither fat tails nor fluctuations in volatility that
are unrelated to the predictability of future returns.
Both models have in common a very strong form of scale-invariance, in which the distribution of
returns over different sampling intervals are identical except for a single, non-random contraction.
This property is at odds with empirical observations. In particular, many financial data sets
become less peaked in the bells and have thinner tails as the sampling interval increases. This does
not necessarily imply that the distribution eventually becomes Gaussian at long enough sampling
intervals. In particular, the multifractal processes introduced in this paper never become Gaussian.
We build up the theory of multifractals in the following sections. Section 3.1 is a general
introduction to multifractality. Section 3.2 presents the binomial measure, which is the simplest
example of a multifractal. Section 3.3 generalizes to the broader case of multiplicative measures.
Section 3.4 extends multifractality from measures to stochastic processes. Section 3.5 presents a
wider definition of multifractal measures based on statistical self-similarity.
3.1 Multifractality
This section provides a general introduction to multifractality. It is aimed at sketching the main
ideas of the theory, while detailed explanations are given in later sections. Multifractal measures
were introduced in Mandelbrot (1972) and have since been applied in the physical sciences to
describe the distribution of energy and matter, e.g. turbulent dissipation, stellar matter, and
minerals. They are new to economics.
This paper uses multifractal measures to model temporal heterogeneity in financial time series.
It also extends multifractality from measures to stochastic processes. For this reason, our presenta-
tion will at times refer to multifractals as either measures or processes. We hope this provides the
reader with the greatest exposure to the mathematical generality of multifractals, without being
unnecessarily confusing.
We previously discussed self-affine processes, which satisfy the simple scaling rule:
d
X(ct) = cH X(t).
8
YALE UNIVERSITY
The theory of multifractals examines the more general relationships:
d
X(ct) = M (c)X(t), (1)
where X and M are independent random functions. Under strict stationarity, arbitrary translations
along the time axis allow extension of (1) to local scaling rules:
d
X(t + c4t) − X(t) = M (c) [X(t + 4t) − X(t)] (2)
for all positive c. The scaling factor M (c) is a random variable, whose distribution does not
depend on the particular instant t. Self-similar processes satisfy (2), with M (c) = cH . To pursue
this analogy, we define the generalized index H(c) = logc M (c), and rewrite the above relation:
d
X(ct) = cH(c) X(t). In contrast to self-similar processes, the index H(c) is a random function of c.
Multifractality thus permits a richer variety of behaviors than is possible under self-affinity.
It also places strong restrictions on the process’s distribution. For instance if c2 /c1 = c3 /c2 and
condition (1) holds, then
X(c2 t) d X(c3 t)
= ,
X(c1 t) X(c2 t)
since both ratios are distributed like M (c2 /c1 ).

d
We also require that the random scaling factor satisfies the property: M (ab) = M1 (a)M2 (b),
where M1 and M2 are independent copies of M . This condition, which is motivated in Section 3.5,
implies the scaling rule:
E (|X(t)|q ) = c(q)tτ (q)+1 , (3)
where τ (q) and c(q) are both deterministic functions of q.
This paper presents the scaling rule (3) as the defining property of multifractal processes. In this
setting, condition (1) only characterizes a particular class of multifractals. Multifractality is thus
defined as a global property of the process’s moments. Our approach could also build on the local
scaling properties of the process’s sample paths11 , in the spirit of equation (2). This alternative
viewpoint is further discussed in the companion paper Calvet, Fisher and Mandelbrot (1997).
We now examine more closely the scaling rule (3). Most of our work concentrates on properties
of the function τ (q), which is called the scaling function. Setting q = 0 in condition (3), we see
11
Local scaling builds on the concept of local Hölder exponent.
9
YALE UNIVERSITY
that all scaling functions have intercept τ (0) = −1. In addition, τ (q) is always concave, as shown
in Section 3.4.
A self-affine process with index H is multifractal, with scaling function τ (q) = Hq − 1. Because
of its linearity, the scaling function is fully determined by a single coefficient, its slope. It is thus
called uniscaling or unifractal. Multifractal (or multiscaling) processes allow more general concave
scaling functions.
Section 4.5 shows how to construct and simulate multifractal processes. We compound a Frac-
tional Brownian Motion BH (t) by the cumulative distribution function θ(t) of a multifractal mea-
sure. The resulting process:
X(t) = BH [θ(t)]
satisfies multiscaling. This construction is the main building block of the MMAR. It requires a
good understanding of multifractal measures, which we now present.
3.2 The Binomial Measure is the Simplest Example of a Multifractal
This section introduces the simplest multifractal, the binomial measure12 on the compact interval
[0, 1]. This is the limit of an elementary iterative procedure called a multiplicative cascade.
Let m0 and m1 be two positive numbers adding up to 1. At stage k = 0, we start the construction
with the uniform probability measure µ0 on [0, 1]. In the step k = 1, the measure µ1 uniformly
spreads mass equal to m0 on the subinterval [0, 1/2] and mass equal to m1 on [1/2, 1]. The density
of µ1 is drawn in Figure 4a for m0 = 0.6.
In step k = 2, the set [0, 1/2] is split into two subintervals, [0, 1/4] and [1/4, 1/2], which respec-
tively receive a fraction m0 and m1 of the total mass µ1 [0, 1/2]. We apply the same procedure to
the dyadic set [1/2, 1] and obtain:
µ2 [0, 1/4] = m0 m0 , µ2 [1/4, 1/2] = m0 m1 ,
µ2 [1/2, 3/4] = m1 m0 , µ2 [3/4, 1] = m1 m1 .
Iteration of this procedure generates an infinite sequence of measures. In step k + 1, we assume

that the measure µk has been defined and construct µk+1 as follows. Consider an interval [t, t+2−k ],
12
The binomial measure is sometimes called the Bernoulli or Besicovitch measure.
10
YALE UNIVERSITY
where the dyadic number t is of the form:
X
k
t = 0.η1 η2 ..ηk = ηi 2−i (4)
i=1
in the counting base b = 2. We uniformly spread a fraction m0 and m1 of the mass µk [t, t + 2−k ] on
the subintervals [t, t + 2−k−1] and [t + 2−k−1, t + 2−k ]. A repetition of this scheme to all subintervals
determines µk+1 . The measure µk+1 is now well-defined. Figure 4b represents the measure µ4
obtained after k = 4 steps of the recursion.
The binomial measure µ is defined as the limit of the sequence (µk ). We now examine some of
its properties. Consider the dyadic interval [t, t + 2−k ], where t = 0.η1 η2 ..ηk in the counting base
b = 2. Let ϕ0 and ϕ1 denote the relative frequencies of 0’s and 1’s in the binary development of t.
The measure of the dyadic interval simplifies to:
µ[t, t + 2−k ] = mkϕ0 kϕ1

0 m1 .
The binomial measure has important characteristics common to many multifractals. It is a contin-
uous but singular probability measure; it thus has no density and no point mass. We also observe
that since m0 + m1 = 1, each stage of the construction preserves the mass of split dyadic intervals.
For this reason, the procedure is called conservative or microcanonical.
This construction can receive several extensions. For instance at each stage of the cascade,
intervals can be split not in 2 but in b > 2 intervals of equal size. Subintervals, indexed from
left to right by β (0 ≤ β ≤ b − 1), receive fractions of the total mass equal to m0 , .., mb−1 . By
P
the conservation of mass, these fractions, also called multipliers, add up to one: mβ = 1. This
defines the class of multinomial measures, which are discussed in Mandelbrot (1989a) and Evertsz
and Mandelbrot (1992).
Another extension randomizes the allocation of mass between subintervals at each step of the
iteration. The multiplier of each subinterval is a discrete random variable Mβ that takes values
m0 , m1 , ..., mb−1 with probabilities p0 , .., pb−1 . The preservation of mass imposes the additivity
P
constraint: Mβ = 1. Figure 4c shows the random density obtained after k = 10 iterations with
parameters b = 2, p = p0 = 0.5 and m0 = 0.6.
11
YALE UNIVERSITY
3.3 Multifractal Measures Generated as Multiplicative Cascades
An extension of the binomial and multinomial measures allows non-negative multipliers Mβ (0 ≤
β ≤ b − 1) that are not necessarily discrete, but can be more general random variables. This
procedure, usually called a multiplicative cascade, helps construct the broader class of multiplicative
measures. To simplify the presentation, we assume that the multipliers are identically distributed,
and denote by M the multiplier M0 .

P
We first impose that mass be preserved at every stage of the construction: Mβ = 1. The
resulting measure is then called conservative or microcanonical. In the first stage of the construction,
the unit interval [0, 1] receives an initial mass equal to 1 and is subdivided into b-adic cells of length
1/b. We index these cells from left to right and allocate for every β the random mass Mβ to the
βth cell.
By a repetition of this scheme, the b-adic cell of length ∆t = b−k , starting at t = 0.η1 ...ηk =
P
ηi b−i , has measure
µ(∆t) = M (η1 )M (η1 , η2 )...M (η1 , ..., ηk ),
and thus [µ(∆t)]q = M (η1 )q M (η1 , η2 )q ...M (η1 , ..., ηk )q for all q ≥ 0. We take the expectation of this
expression and obtain the scaling rule:
E [µ(∆t)q ] = [E (M q )]k , (5)
since the multipliers are independent.
Modifying the previous construction, we now choose that the multipliers Mβ be statistically
P
independent. Each iteration only conserves mass “on average” in the sense that E ( Mβ ) = 1
or E M = 1/b. The corresponding measure is then called canonical. Its total mass, denoted Ω, is
generally random13 , and the mass of a b-adic cell takes the form:
µ(∆t) = Ω(η1 , ..., ηk )M (η1 )M (η1 , η2 )...M (η1 , ..., ηk ).
We note that Ω(η1 , ..., ηk ) has the same distribution as Ω. The measure µ thus satisfies the scaling
relationship:
E [µ(∆t)q ] = E (Ωq ) [E (M q )]k , (6)

13
The random variable Ω has interesting distributional and tail properties that are discussed in Mandelbrot (1989a).
12
YALE UNIVERSITY
which characterizes multifractals. In the companion empirical paper, we argue that this condition
on population expectations can be extended to sample sums of data, allowing us to test 1) for
scaling behavior of the moments, and 2) to distinguish between unifractal and multifractal scaling
behavior.
This section presented examples of multifractal measures that were constructed as the limit of
multiplicative cascades. We did not give a general definition of multifractality. A broader approach,
based on the geometric concept of self-similarity, provides better intuition of multifractal measures
and is presented in section 3.5. This larger setting is slightly more complicated, and can be skipped
in a first reading of the paper.
3.4 Multifractal Processes
We now extend multifractality from measures to stochastic processes. This extension is new to
this paper and Mandelbrot (1997). We find it convenient to define multifractal processes in terms
of moments, because this has direct graphical and testable implications. In Calvet, Fisher and
Mandelbrot (1997), we focus on the local scaling properties of multifractal processes. This alter-
native leads to a more elegant mathematical presentation, and for some, perhaps a more intuitive
understanding of multifractality. For now, we concentrate on the simpler idea of scaling in the
moments of the process’s increments.
Definition 2 A stochastic process {X(t)} is called multifractal if it has stationary increments and
satisfies:
E (|X(t)|q ) = c(q)tτ (q)+1 , for all t ∈ T , q ∈ Q, (7)
where T and Q are intervals on the real line, τ (q) and c(q) are functions with domain Q. Moreover,
we assume that T and Q have positive lengths, and that 0 ∈ T , [0, 1] ⊆ Q.
A multifractal process is thus globally scaling, in the sense that its moments satisfy the scaling
relationship (7). The function τ (q) is called the scaling function of the multifractal process. Setting
q = 0 in condition (7), we see that all scaling functions have the same intercept τ (0) = −1.
Self-affine processes are multifractal, as is now shown. A self-affine process {X(t), t ≥ 0},
d
with self-affinity index H, satisfies X(t) = tH X(1), and therefore E (|X(t)|q ) = tHq E (|X(1)|q ) .
Condition (7) thus holds, with:
τ (q) = Hq − 1 and c(q) = E (|X(1)|q ) .
13
YALE UNIVERSITY
In the special case of self-affine processes, the scaling function τ (q) is linear and fully determined
by its index H. More generally, linear scaling functions τ (q) are determined by a unique parameter,
their slope. For this reason, multifractal processes with linear τ (q) are called uniscaling or unifractal.
In this paper, we focus instead on multifractal processes with non-linear functions τ (q). Such
processes are also called multiscaling.
The concavity of τ (q) is easy to derive from condition (7). Consider two exponents q1 , q2 , and
two positive weights w1 , w2 adding up to one. Hölder’s inequality implies that:
E (|X(t)|q ) ≤ [E (|X(t)|q1 )]w1 [E (|X(t)|q2 )]w2 ,
where q = w1 q1 + w2 q2 . Taking logarithms and using (7), we obtain:
ln c(q) + τ (q) ln t ≤ [w1 τ (q1 ) + w2 τ (q2 )] ln t + [w1 ln c(q1 ) + w2 ln c(q2 )]. (8)
We divide by ln t < 0, and let t go to zero:
τ (q) ≥ w1 τ (q1 ) + w2 τ (q2 ), (9)
which establishes the concavity of τ .
In fact this proof contains additional information on multifractal processes. Assuming that
relation (7) holds for t ∈ [0, ∞), we divide inequality (8) by ln t > 0 and let t go to infinity. We
obtain the reverse of inequality (9), and conclude that τ (q) is linear. Thus multiscaling can only
hold for bounded time intervals T . We can reinterpret this results as follows. Processes defined
on unbounded intervals can only be multifractal over bounded ranges of time. They must contain
what physicists call crossovers, i.e. transitions in their scaling properties. Mandelbrot (1997) also
discusses this result. This technical difficulty has little consequence for financial modeling, since
multifractal processes can be defined on arbitrarily large time intervals.
We now describe a large class of multiscaling processes, which is inspired by the previous
discussion on self-similar random measures. Consider a process {X(t)}, and assume the existence
of an independent process {M (c)} that satisfies:
d
X(ct) = M (c)X(t), for all t, 0 < c ≤ 1.
and
14
YALE UNIVERSITY
Property 1. If 0 < a, b ≤ 1, the process M takes positive values and satisfies:
d
M (ab) = M1 (a)M2 (b), (10)
where M1 and M2 are two independent copies of M .
Property 1 implies that E [M (ab)q ] = E [M (a)q ] E [M (b)q ] for all q ≥ 0. When these moments
are finite, the process M satisfies the scaling relationship:
E [M (c)q ] = cτ (q)+1 ,
and the process {X(t)} is multifractal.
This section has defined multifractality as a scaling property of the process’s moments. This
presentation has the advantage of having directly testable implications, but somewhat lacks intuitive
content. For this reason, we now present an alternative interpretation of multifractality based on
the concept of self-similarity.
3.5 Self-Similar Random Measures
[Please note: This section is not necessary to understand the Multifractal Model of Asset Returns,
but generalizes some of the previous presentation.]
We now present a more general approach to multifractality based on the statistical self-similarity
of random measures. To simplify the exposition, we only consider the case of random measures
defined on an interval X of the real line. An extension of self-similarity to higher dimensions can
be found in Mandelbrot (1989a.)
A random measure µ on the interval X is analogous to a random variable. It is a mapping
defined on a probability space, and valued in the class of all the measures on X. Moreover given a
fixed interval I ⊆ X, the mass µ(I) is a random variable.

We now present the conditions on the random measure that define statistical self-similarity.
This will help our discussion of the multifractal market model, in which trading time is viewed
as the cumulative distribution function (c.d.f.) of a self-similar random measure. In a general
Euclidean space, we call similitude the compound of a translation, a homothetic transformation,
and a rotation. The class of similitudes on the real line reduces to the class of affine transformations,
15
YALE UNIVERSITY
and is denoted by S. We first impose that conditional measures be statistically invariant under
similitudes.
Assumption 1. For any S ∈ S, for any intervals I1 ⊆ I2 , the ratios
µ(SI1 ) µ(I1 )
and
µ(SI2 ) µ(I2 )
are identically distributed whenever I1 , I2 , SI1 , SI2 ⊆ X.
Successive iterations of a multiplicative cascade are statistically independent. This property
generalizes as follows.
Assumption 2. For all non-decreasing sequence of compact intervals I1 ⊆ ... ⊆ In contained in
X, the random variables
µ(I1 ) µ(In−1 )
, ...,
µ(I2 ) µ(In )
are statistically independent.
This leads to the following
Definition 3 A random measure satisfying Assumptions [1] and [2] is called self-similar.
When the interval X is of the form [0, T ], 0 < T ≤ ∞, Assumption [1] implies the existence of
a positive random process M (c) independent of µ that satisfies:
d
µ[0, ct] = M (c)µ[0, t] whenever 0 < t ≤ T, 0 < c ≤ 1, (11)
We note in particular that M (1) = 1. Given two coefficients a, b ≤ 1, we can write:
µ[0, abt] µ[0, abt] µ[0, at]

= .
µ[0, t] µ[0, at] µ[0, t]
By Assumption [2], the two ratios on the right-hand side are statistically independent, and the
process M satisfies Property 1.14
The multiplicative measures of the previous sections do not exactly satisfy self-similarity. In
these examples, the multiscaling relation (11) only holds for certain values of c and t. For instance
14
Property 1 was defined in section 3.4.
16
YALE UNIVERSITY
in the case of the binomial measure, relation (11) only applies when c and t are dyadic numbers.
When the relation only holds on a dense set of t’s, we say that µ is grid-bound self-similar. Such
is the case of the binomial measure. By contrast, the stronger definition given in this section
characterize grid-free self-similar measures.
We can now discuss the moments of the measure. Assume without loss of generality that
X = [0, 1], and consider the random variables µ(a) ≡ µ[0, a]. As in the previous section, Property
1 implies that:
E M (a)q = aτ (q)+1 ,
and therefore:
E [µ(a)q ] = E[µ(1)q ]aτ (q)+1 . (12)
The scaling relation is thus a direct consequence of statistical self-similarity. This provides some
justification for the definition of multifractal processes in terms of moments.
We briefly examine the properties of τ (q). As in previous sections, setting q = 0 in (12) yields
τ (0) = −1. For q = 1, we partition the unit interval [0, 1] into n subintervals of equal length, and
write that the mass of the subintervals add up to µ(1). Taking expectations yields E [µ(1/n)] =
E [µ(1)]/n, and the scaling function thus satisfies τ (1) = 0. By Hölder’s inequality, we also know
that τ (q) is concave.
This framework helps model temporal heterogeneity in financial time series. Large price changes
tend to be concentrated in time, and it is natural to assume that the distribution of volatility across
time be statistically self-similar. We formalize this intuition in the MMAR.
4 The Multifractal Model of Asset Returns
This section uses the multifractal processes introduced in Section 3 to build a new financial model,
the Multifractal Model of Asset Returns (MMAR). In this framework, the price of a financial asset
is viewed as a multiscaling process with long memory and long tails. Fluctuations in “volatility”
are introduced in the MMAR by a random trading time, generated as the c.d.f. of a random
multifractal measure. This construction is new to both mathematics and finance.
17
YALE UNIVERSITY
We present the multifractal model in the following sections. Section 4.1 introduces trading time
and compounding. Section 4.2 defines the MMAR. Section 4.3 compares the MMAR to earlier
models. Section 4.4 presents simulation results.
4.1 Trading Time and Compound Processes
Trading time is the key concept facilitating the application of multifractals to financial markets.
We introduce the following

Definition 4 Let {B(t)} be a stochastic process, and θ(t) an increasing function of t. The process
X(t) ≡ B[θ(t)]
is called a compound process. The index t denotes clock time, and θ(t) is called trading time or the
time deformation process.
A special form of compound process is subordination, as developed by Bochner (1955), and
applied to financial markets by Mandelbrot and Taylor (1968). Subordination originally developed
as part of the theory of Markov processes, and requires that θ(t) has independent increments.15
In economics, the concept of subordination has evolved differently, and now encompasses any
generic time deformation process.16 For example, Dacorogna et al. (1992, 1993), use trading time
to model time-of-day and day-of-week seasonality. Ghysels, Gouriéroux, and Jasiak (1996), model
a stochastic trading time with volatility conditioned on measures of current market activity.
The MMAR posits a trading time that is the c.d.f. of a multifractal measure. Thus, trading
time will be both highly variable and contain long memory. Both of these characteristics will be
passed on to the price process through compounding. The main significance of compounding is that
it allows direct modelling of a processes’ variability without affecting the direction of increments
or their correlations.
4.2 The Multifractal Model of Asset Returns
This section presents a new model for the price of a financial asset {P (t); 0 ≤ t ≤ T }. We introduce
the notation:
X(t) = ln P (t) − ln P (0),

15
See Feller (1968) p. 355. More current references to subordination in the mathematics and statistics literature
include Rogers and Williams (1987) and Bertoin (1996).
16
The different usage of subordination in economics can be traced back at least to Clark (1973). This paper
develops all of its theory using Markov assumptions, and in empirical work proposes trading volume as a directing
process.
18
YALE UNIVERSITY
and assume the following:
Assumption 1. X(t) is a compound process:
X(t) ≡ BH [θ(t)]
where BH (t) is a fractional Brownian Motion with self-affinity index H, and θ(t) is a stochas-
tic trading time.
Assumption 2. The trading time θ(t) is the c.d.f. of a multifractal measure defined on [0, T ].
That is, θ(t) is a multifractal process with continuous, non-decreasing paths, and stationary
increments.
Assumption 3. {BH (t)} and {θ(t)} are independent.
The trading time θ(t) plays a crucial role in the MMAR. We first note that θ(0) = 0 almost
surely since by definition X(0) = 0. Assumption 2 imposes that θ(t) be the cumulative distribution
function of a self-similar random measure, such as a binomial or a multiplicative measure. Trading
time θ(t) presumably causes the price X(t) to be multifractal, and we expect the scaling functions
τθ (q) and τX (q) to be closely related. This intuition leads to the following
Theorem 5 Under Assumptions [1] − [3], the process X(t) is multifractal, with scaling function
τX (q) ≡ τθ (Hq) and stationary increments.
Proof: See Appendix.
The above construction generates a large class of multifractal processes. Although a complete
theory is not currently available, we now discuss some of the most important properties of the
MMAR.
4.3 Properties of the MMAR
We first examine tail properties. The multiscaling relation (7) imposes that if E |X(t)|q is finite for
some instant t, then it is finite for all t. This justifies dropping the time index when discussing the
moments of multifractal processes. By Theorem 4.2, the q-th moment of X exists if (and only if)
the process θ has a moment of order Hq. The trading time thus controls the moments of the price
X(t).
19
YALE UNIVERSITY
By assumption, the trading time is generated by a self-similar measure µ. Its moments have
very different properties depending upon whether µ is microcanonical or canonical. By definition,
microcanonical random measures have a fixed mass on [0, T ] in the construction. The corresponding
θ(t) are therefore bounded, and the compound process X(t) has finite moments of all (non-negative)
order. Microcanonical measures thus generate “mild” processes with relatively thin tails. On the
other hand, canonical measures permit financial models that have diverging moments. Section 3.3
shows that the total mass of a canonical measure is a random variable Ω, which fully determines the
tail behavior of prices. Mandelbrot (1972) conjectures and Guivarc’h (1987) proves that Ω generally
has Paretian tails and infinite moments.17 The corresponding process X(t) is then “wild”. Overall,
the MMAR has enough flexibility to accommodate a wide variety of tail behaviors.
We now study correlation in the process’s increments. For a fixed ∆t > 0 and any process Z,
we define:
Z(t, ∆t) = Z(t + ∆t) − Z(t),
and the covariance function:
γZ (t) = Cov [Z(a, ∆t), Z(a + t, ∆t)] .
Since Z(t) has stationary increments, we know that γZ (t) does not depend on the choice of a . It
is natural to first consider the special case H = 1/2, in which BH (t) is a Brownian Motion. We
prove the following
Theorem 6 If BH (t) is a Brownian motion without drift, the following properties hold:
1. If E (θ 1/2 ) is finite, then {X(t)} is a martingale with respect to its natural filtration.
2. If E θ is finite, the increments of X(t) are uncorrelated, that is γX (t) = 0 for all t ≥ ∆t.
Theorem 4.3 shows that when H = 1/2, the MMAR generates a price process that has a white
17
As a consequence, there exists a critical exponent qcrit (θ) > 1. The moment E θq is finite when 0 ≤ q < qcrit (θ),
and infinite when q ≥ qcrit (θ). Moreover, the scaling function τθ (q) is negative when 0 < q < 1 and positive when
1 < q < qcrit (θ). Mandelbrot (1990) provides additional discussion of this topic.
20
YALE UNIVERSITY
spectrum.18 This result builds on the martingale property of the Brownian motion, and does not
extend to the case H 6= 1/2. We can in fact prove the following
Theorem 7 If E (θ 2H ) is finite, the autocovariance function of the price process X(t) satisfies for
all t ≥ ∆t :
γX (t) = K {(t + ∆t)m + (t − ∆t)m − 2tm } (13)
where m = τθ (2H) + 1 and K = cθ (2H)V ar[BH (1)]/2. It is positive when H > 1/2, and negative
when H < 1/2.
When H > 1/2, the process BH (t) has long memory, and price increments are positively cor-
related. The definition of long memory for the price process, which may only be defined on a
bounded time range, is more delicate. This introductory paper informally defines long memory by
the following geometric property: the longest apparent cycle has approximately the same length as
the interval of definition. In this sense, the c.d.f. of a multiplicative cascade has long memory. By
Theorem 4.4, it seems safe to conjecture that the price process has long memory when H > 1/2.
The MMAR thus allows for a wide variety of autocorrelation structures.
We finally examine dependence in the absolute values of returns, which will indicate whether the
MMAR displays persistence in volatility. For any stochastic process Z with stationary increments,
it is convenient to define:
δZ (t, q) = Cov(|Z(a, ∆t)|q , |Z(a + t, ∆t)|q ).
We now prove the following
Theorem 8 If H ≥ 1/2 and E (θ Hq ) is finite, the compound process satisfies:
δX (t, q) ≥ δθ (t, Hq) [E |BH (1)|q ]2 (14)
for all non-negative q and t ≥ ∆t. Moreover, this result holds as an equality when H = 1/2.
18
Theorem 4.3 only shows that ln P (t) is a martingale. By Jensen’s inequality, the price P (t) is then a submartingale,
but not a martingale. A very unstable concept, the martingale property is only discussed to illustrate the flexibility
of the model. Moreover, a detailed study of the relation between the MMAR and market efficiency is beyond the
scope of this paper.
21
YALE UNIVERSITY
Theorem 4.5 indicates that the price process has long memory in the absolute value of its
increments. In particular when H = 1/2, the price process displays both uncorrelated increments
and persistence in volatility. Thus while the MMAR construction is certainly not familiar in finance,
it has quite natural and appealing implications for empirical applications. The MMAR allows for
long tails, correlated “volatilities”, and either unpredictability or long memory in returns, and thus
combines the properties of many earlier models.
4.4 Comparison with Earlier Models
This section discusses similarities and differences between the MMAR and earlier models. Con-
clusion 1 of Theorem 4.3 guarantees that, when {X(t)} is a Brownian Motion without drift, the
direction of future returns is not predictable from knowledge of past prices. At the same time,
Theorem 4.5 guarantees long memory in the absolute value of returns. That is, there is not only
temporal heterogeneity in the size of returns, but this temporal heterogeneity is present on any
scale at which we choose to look at the data - whether daily, weekly, monthly, or yearly returns.
This result may surprise readers who have learned to associate (from the FBM) long memory
with the predictability of market returns. In fact, the MMAR is not the first to combine long
memory with a martingale property. The FIGARCH model of Baillie, Bollerslev and Mikkelsen
(1996) also has this feature. The most important distinguishing feature of the multifractal model
is that returns are scale-consistent, while FIGARCH is not.
To place the multifractal market model more concretely within the existing literature on the
distribution of financial returns, we present the following tables.
22
YALE UNIVERSITY
Table 1: Typical Characteristics for Models of Financial Returns
Volatility Clustering Volatility Clustering PROPERTIES
Implies Consistent with
Predictable Price Martingale Price
ARMA∗ GARCH no long memory
scale-inconsistent
ARFIMA∗ FIGARCH long memory
scale-inconsistent
FBM MMAR long memory
scale-consistent
∗ Asymptotically, ARMA scales like Brownian Motion, and ARFIMA like FBM.
Table 2: Typical Covariance Characteristics of Models of Returns
Cov(|Xt+s+4t − Xt+s |, |Xt+4t − Xt |) Cov(Xt+s+4t − Xt+s , Xt+4t − Xt )
ARMA weak weak
GARCH weak zero
ARFIMA strong strong
FIGARCH strong zero
FBM strong strong
MMAR strong zero or strong
The distinguishing features of the multifractal market model are thus:
1. Long memory in volatility
2. Compatibility with the martingale property of returns
3. Scale-consistency
4. Multiscaling
Properties 1 and 2 are given by Theorems 4.5 and 4.3. Properties 3 and 4 are consequences of
the definition of multifractal processes. They respectively correspond to the time-invariance and
nonlinearity of the scaling function τ (q).
23
YALE UNIVERSITY
4.5 Simulation of Multifractal Processes
We can generate a wide range of continuous multifractal processes with sequences of increments
that are indistinguishable from real financial data at all time scales.
Example 9 Figures 5a and 5b show two simulated sample paths of multifractal processes. Figures
6a and 6b show their respective first differences. In these simulations, {B(t)} is a standardized
Brownian Motion. {θ(t)} is the c.d.f. of a randomized binomial measure with multiplier m0 = .6.
The most striking aspect of these simulations is their strong resemblance in character to the
exchange rate data in Figure 2. In contrast to the GARCH simulation in Figure 1, these simula-
tions display marked temporal heterogeneity at all time scales. In addition, we observe crashes,
upcrashes, slow but persistent growth and decline, and a variety of behavior interesting to technical
analysts. This is quite atypical of the Brownian Motion and GARCH processes.
Other factors are of far greater importance to financial economists. Foremost among these
are the testable implications of multifractality. Moreover, the clearly non-Gaussian behavior of
multifractal processes has tremendous implications for the behavior of risk-averse investors and the
pricing of financial derivatives.
5 Conclusion
The Multifractal Model of Asset Returns, which is new to this paper and Mandelbrot (1997),
incorporates important regularities observed in financial time series including long tails and long
memory. Multifractality is defined by a set of restrictions on the process’s moments as the time scale
of observations changes. It is integrated in the model through trading time, a random distortion of
clock time that accounts for changes in volatility.
The properties of the MMAR include multifractality, scale-consistency and long memory in
volatility. The predictability of (log) prices is not theoretically specified, since the model has
enough flexibility to satisfy the martingale property in some cases and long memory in its increments
otherwise. The MMAR is thus a promising alternative to ARCH-type models. Like FIGARCH,
the MMAR incorporates long memory in volatility. In addition, the MMAR allows the possibility
that returns are uncorrelated, but does not require it. This is an important property for researchers
interested in issues of market efficiency. The main advantage of the MMAR over alternatives like
FIGARCH is the property of scale-consistency. Because of this property, aggregation characteristics
24
YALE UNIVERSITY
of the data (otherwise thought of as the information contained at different sampling frequencies)
can be used to test and identify the model.
The main disadvantage of the MMAR is the dearth of applicable statistical methods. We
propose that new econometric methods are needed for models which are both time-invariant and
scale-invariant. A demonstration of this type of method appears in the companion paper, Fisher,
Calvet, and Mandelbrot (1997).
25
YALE UNIVERSITY
6 Appendix
6.1 Proof of Theorem 4.2
The argument builds on the iterated expectation:
E (|X(t)|q ) = E [E (|X(t)|q |θ(t) = u )]
Since the trading time and the self-affine process {BH (t)} are independent, conditioning on θ(t)
yields:
E (|X(t)|q |θ(t) = u ) = E [|BH (u)|q |θ(t) = u ]
= θ(t)Hq E [|BH (1)|q ] ,
and thus

E [|X(t)|q ] = E θ(t)Hq E [|BH (1)|q ]
The process X(t) therefore satisfies multiscaling relation (7), with τX (q) ≡ τθ (Hq) and cX (q) ≡
cθ (Hq) E [|BH (1)|q ] .
1. Let Ft and Ft0 respectively denote the natural filtrations of {X(t)} and {X(t), θ(t)}. We
compute E { X(t + T )| Ft } as the iterated expectation:

E E BH [θ(t + T )]| Ft0 , θ(t + T ) = u Ft
For any t, T and u ≥ t, the independence of BH and θ implies that:

E BH [θ(t + T )]| Ft0 , θ(t + T ) = u = E BH (u)|Ft0
= BH [θ(t)]
since {BH (t)} is a martingale in this case. We now infer that
E [X(t + T )|Ft ] = X(t).
2. This property, which holds for all square integrable martingales, is a direct consequence of
the fact that E [X(t0 + ∆t) − X(t0 )|Ft ] = 0 when t + ∆t ≤ t0 .
26
YALE UNIVERSITY
Consider the conditional expectation:
E { X(0, ∆t)X(t, ∆t)| θ(∆t) = u1 , θ(t) = u2 , θ(t + ∆t) = u3 } . (15)
Since BH (t) and θ(t) are independent processes, this expression simplifies to:
E {BH (u1 )[BH (u3 ) − BH (u2 )]} .
or19
1n o
|u3 |2H − |u2 |2H + |u2 − u1 |2H − |u3 − u1 |2H V ar[BH (1)],
2
which we rewrite in terms of trading time:
V ar[BH (1)] n o
|θ(t + ∆t)|2H − |θ(t)|2H + |θ(t) − θ(∆t)|2H − |θ(t + ∆t) − θ(∆t)|2H .
2
Taking the expectation yields (13).

τθ (2H)+1
The sign of γX (t) is determined by the concavity of the function f (x) = x , i.e. by the
sign of τθ (2H). We know that τθ (1) = 0, so that for all q satisfying 0 ≤ q < qcrit (θ), the scaling
function τθ (q) has the same sign as q − 1. When H = 1/2, the autocovariance equals zero since
τθ (2H) = 0. When H > 1/2, f (x) is convex and the function γX (t) is positive. Conversely γX (t)
is negative for H < 1/2.
The argument builds on the properties of the conditional expectation:
E { |X(0, ∆t)X(t, ∆t)|q | θ(∆t) = u1 , θ(t) = u2 , θ(t + ∆t) = u3 } . (16)
We start by rewriting this quantity:
E {|BH (u1 )[BH (u3 ) − BH (u2 )]|q }
since BH (t) and θ(t) are independent processes.

19
For FBM, we know that:

E [BH (u1 )BH (u2 )] = |u1 |2H + |u2 |2H − |u1 − u2 |2H V ar[BH (1)]/2.
27
YALE UNIVERSITY
We assumed that the FBM has either independent (H = 1/2) or positively correlated (H > 1/2)
increments. Therefore expression (16) is equal to, or bounded below by:
2
E [|BH (u1 )|q ] E [|BH (u3 ) − BH (u2 )|q ] = |u1 |Hq |u3 − u2 |Hq [E |BH (1)|q ] .
In terms of trading time, this lower bound can be rewritten:
|θ(∆t)|Hq |θ(t, ∆t)|Hq [E |BH (1)|q ]2 . (17)
Taking the expectation of (16) and (17), we infer that:
2
E [ |X(0, ∆t)X(t, ∆t)|q ] ≥ E [ |θ(0, ∆t)θ(t, ∆t)|q ] [ E |BH (1)|q ] .
The rest of the proof is straightforward.
28
YALE UNIVERSITY
References
[1] Bachelier, L. (1900), Théorie de la Spéculation, Annales de l’Ecole Normale Supérieure 3,
Paris: Gauthier Villars. English translation in Cootner (1964)
[2] Baillie, R. T. (1996), Long Memory Processes and Fractional Integration in Econometrics,
Journal of Econometrics 73, 5-59
[3] Baillie, R. T., Bollerslev, T., and Mikkelsen, H. O. (1996), Fractionally Integrated Generalized
Autoregressive Conditional Heteroskedasticity, Journal of Econometrics 74, 3-30
[4] Belkacem, L., Lévy-Véhel, J., and Walter, C. (1995), Generalized Market Equilibrium: “Sta-
ble” CAPM, Working Paper presented at the AFFI-International Conference of Finance 1995
held in Bordeaux, France
[5] Bertoin, J. (1996), Lévy Processes, Cambridge, UK: Cambridge University Press
[6] Bochner, S. (1955), Harmonic Analysis and the Theory of Probability, Berkeley: University of
California Press
[7] Bollerslev, T. (1986), Generalized Autoregressive Conditional Heteroskedasticity, Journal of

Econometrics 31, 307-327
[8] Bollerslev, T., Engle, R. F., and Nelson, D. B. (1994), ARCH Models, ch 49 in: R. Engle and
D. McFadden eds., Handbook of Econometrics, Vol.4, Amsterdam: North-Holland
[9] Calvet, L., Fisher, A., and Mandelbrot, B. B. (1997), Large Deviation Theory and the Distri-
bution of Price Changes, Yale University, Working Paper
[10] Clark, P. K. (1973), A Subordinated Stochastic Process Model with Finite Variance for Spec-
ulative Prices, Econometrica 41, 135-156
[11] Cootner, P. (1964), The Random Character of Stock Market Prices, Cambridge, MA: MIT
Press
[12] Dacorogna, M. M., Müller, U. A., Nagler, R. J., Olsen, R. B., and Pictet, O. V. (1993), A
Geographical Model for the Daily and Weekly Seasonal Volatility in the Foreign Exchange
Market, Journal of International Money and Finance 12, 413-438
[13] Ding, Z., Granger, C. W. J., and Engle, R. F. (1993), A Long Memory Property of Stock
Returns and a New Model, Journal of Empirical Finance 1, 83-106
[14] Drost, F. C., and Nijman, T. E. (1993), Temporal Aggregation of GARCH Processes, Econo-
metrica 61, 909-927
[15] Drost, F. C., and Werker, B. J. M. (1996), Closing the GARCH Gap: Continuous Time
GARCH Modeling, Journal of Econometrics 74, 31-57
[16] Engle, R. F. (1982), Autoregressive Conditional Heteroscedasticity with Estimates of the Vari-
ance of United Kingdom Inflation, Econometrica 50, 987-1007
[17] Engle, R. F., Lilien, D. M., and Robbins, R. P. (1987), Estimating Time Varying Risk Premia
in the Term Structure: the ARCH-M Model, Econometrica 55, 391-408
29
YALE UNIVERSITY
[18] Evertsz, C. J. G., and Mandelbrot, B. B. (1992), Multifractal Measures, in: Peitgen, H. O.,
Jürgens, H., and Saupe, D. (1992), Chaos and Fractals: New Frontiers of Science, 921-953,
New York: Springer Verlag
[19] Fama, E. F. (1963), Mandelbrot and the Stable Paretian Hypothesis, Journal of Business 36,
420-429
[20] Feller, W. (1968), An Introduction to Probability Theory and Its Applications, Volume I, New
York: Wiley and Sons
[21] Fisher, A., Calvet, L., and Mandelbrot, B. B. (1997), Multifractality of Deutschmark/US
Dollar Exchange Rates, Yale University, Working Paper
[22] Frisch, U., and Parisi, G. (1985), Fully Developed Turbulence and Intermittency, in: M. Ghil
ed., Turbulence and Predictability in Geophysical Fluid Dynamics and Climate Dynamics, 84-
88, Amsterdam: North-Holland
[23] Ghysels, E., Gouriéroux, C., and Jasiak, J. (1995), Market Time and Asset Price Movements:
Theory and Estimation, CIRANO and CREST Discussion Paper
[24] Ghysels, E., Gouriéroux, C., and Jasiak, J. (1996), Trading Patterns, Time Deformation and
Stochastic Volatility in Foreign Exchange Markets, CREST Working Paper n ◦ 9655
[25] Granger, C. W. J., and Joyeux, R. (1980), An Introduction to Long Memory Time Series
Models and Fractional Differencing, Journal of Time Series Analysis 1, 15-29
[26] Guivarc’h, Y. (1987), Remarques sur les Solutions d’une Equation Fonctionnelle Non Linéare
de Benoit Mandelbrot, Comptes Rendus (Paris) 3051, 139
[27] Halsey, T. C., Jensen, M. H., Kadanoff, L. P., Procaccia, I., and Shraiman, B. I. (1986), Frac-
tal Measures and their Singularities: The Characterization of Strange Sets, Physical Review
Letters A 33, 1141
[28] Hosking, J. R. M. (1981), Fractional Differencing, Biometrika 68, 165-176
[29] Koedjik, K. G., and Kool, C. J. M. (1992), Tail Estimates of East European Exchange Rates,
Journal of Business and Economic Statistics 10, 83-96
[30] Lévy, P. (1925), Calcul des Probabilités, Paris: Gauthier-Villars
[31] Lévy, P. (1937), Théorie de l’Addition des Variables Aléatoires, Paris: Gauthier-Villars
[32] Mandelbrot, B. B. (1963), The Variation of Certain Speculative Prices, Journal of Business
36, 394-419
[33] Mandelbrot, B. B. (1967), The Variation of Some Other Speculative Prices, Journal of Business
40, 393-413
[34] Mandelbrot, B. B. (1972), Possible Refinements of the Lognormal Hypothesis Concerning the
Distribution of Energy Dissipation in Intermitent Turbulence, in: M. Rosenblatt and C. Van
Atta eds., Statistical Models and Turbulence, New York: Springer Verlag
[35] Mandelbrot, B. B. (1974), Intermittent Turbulence in Self Similar Cascades; Divergence of

High Moments and Dimension of the Carrier, Journal of Fluid Mechanics 62
30
YALE UNIVERSITY
[36] Mandelbrot, B. B. (1977), Fractals: Form, Chance and Dimension, San Francisco: Freeman
[37] Mandelbrot, B. B. (1982), The Fractal Geometry of Nature, New York: Freeman
[38] Mandelbrot, B. B. (1989a), Multifractal Measures, Especially for the Geophysicist, Pure and
Applied Geophysics 131, 5-42
[39] Mandelbrot, B. B. (1989b), Examples of Multinomial Multifractal Measures that Have Neg-
ative Latent Values for the Dimension f (α), in: L. Pietronero ed., Fractals’Physical Origins
and Properties, 3-29, New York: Plenum
[40] Mandelbrot, B. B. (1990), Limit Lognormal Multifractal Measures, in: E. A. Gotsman et al.
eds., Frontiers of Physics: Landau Memorial Conference, 309-340, New York: Pergamon
[41] Mandelbrot, B. B. (1997), Fractals and Scaling in Finance: Discontinuity, Concentration,

Risk, New York: Springer Verlag
[42] Mandelbrot, B. B., and Ness, J. W. van (1968), Fractional Brownian Motion, Fractional Noises
and Application, SIAM Review 10, 422-437
[43] Mandelbrot, B. B. and Taylor, H. W. (1967), On the Distribution of Stock Price Differences,
Operations Research 15, 1057-1062
[44] Mills, T. C. (1993), The Econometric Modelling of Financial Time Series, Cambridge: Cam-
bridge University Press
[45] Müller, U. A., Dacorogna, M. M., Davé, R. D., Pictet, O. V., Olsen, R. B., and Ward, J.
R. (1995), Fractals and Intrinsic Time: A Challenge to Econometricians, Discussion Paper
Presented at the 1993 International Conference of the Applied Econometrics Association held
in Luxembourg
[46] Nelson, D. B. (1991), Conditional Heteroskedasticity in Asset Returns: A New Approach,

Econometrica 59, 347-370
[47] Phillips, P. C. B., McFarland, J. W., and McMahon, P. C. (1994), Robust Tests of Forward
Exchange Market Efficiency with Empirical Evidence from the 1920’s, Journal of Applied
Econometrics 11, 1-22
[48] Rogers, L. C. G., and Williams, D. (1987), Diffusions, Markov Processes and Martingales,
New York: Wiley
[49] Samorodnitsky, G., and Taqqu, M. S. (1994), Stable Non-Gaussian Random Processes, New
York: Chapman and Hall
[50] Taqqu, M. S. (1975), Weak Convergence to Fractional Brownian Motion and to the Rosenblatt
Process, Z. Wahrscheinlichkeitstheorie verw. Gebiete 31, 287-302
[51] Taylor, S. (1986), Modelling Financial Time Series, New York: Wiley
31
YALE UNIVERSITY
YALE UNIVERSITY
YALE UNIVERSITY
YALE UNIVERSITY
YALE UNIVERSITY
YALE UNIVERSITY

A Multi Fractal Model of Asset Returns

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Multi Fractal Model of Asset Returns

Uploaded by

Copyright:

Available Formats

COWLES FOUNDATION FOR RESEARCH IN ECONOMICS

A Multifractal Model of Asset Returns

Adlai Fisher and Laurent Calvet†

Cowles Foundation Discussion Paper #1164

This Draft: September 15, 1997

THIS VERSION TEXT ONLY

now well-known in finance. First, the MMAR contains long-tails, as in Mandelbrot

(1963), which focused on Lévy-stable distributions. In contrast to Mandelbrot (1963),

long-dependence, the characteristic feature of fractional Brownian Motion (FBM), in-

increments themselves can be uncorrelated. As such, the MMAR is an alternative to

model is multiscaling of the return distribution’s moments under time-rescalings. We

multifractal model implies certain empirical regularities, which are investigated in a

Keywords: Multifractal Model of Asset Returns, Compound Stochastic Process, Sub-

scaling, Self-Similarity, Self-Affinity

much fatter than predicted by the Gaussian distribution.

tions in returns are specified as

where ut are i.i.d., and

The common strand in GARCH-type representations is a conditional distribution of returns

data, and mitigates the problem of fat tails.1

a well documented stylized fact of financial time-series.2

of mild conditional non-stationarity characteristic of GARCH representations with low values of α

number of lags) periods.

the presence of temporal dependence of varying frequencies – a phenomenon which non-integrated

is a strong trade-off, because of the stability requirements of GARCH representations (e.g.

(weak-memory) GARCH-type representations do not capture.

the number of observations.

simultaneously with long memory in the absolute value of returns.

no convergence to Brownian behavior over very long sampling intervals.

to the model when choosing the time-scale of the data.

with a weak GARCH(1, 1) representation, there is a unique correspondence between values of

this correspondence should be considered jump-diffusions.

aspect of multifractality - a scaling property in moments of the process or measure. In fact, we

and discuss interesting properties of the MMAR from an economic perspective.

more consistent with the MMAR hypothesis.

2 Roots of the MMAR

imply an infinite variance of returns over discrete sampling intervals.

the possibility that returns themselves are white.

2.2 Self-Affine Processes

Definition 1 Granted X(0) = 0, a random process {X(t)} that satisfies:

We call H the self-affinity index, or scaling exponent, of X(t).

under dilations, but also rotations.

when 1/2 < H < 1. Persistent FBM have long memory.

3 Multifractal Measures and Processes

are unrelated to the predictability of future returns.

wider definition of multifractal measures based on statistical self-similarity.

minerals. They are new to economics.

The theory of multifractals examines the more general relationships:

condition (1) holds, then

since both ratios are distributed like M (c2 /c1 ).

implies the scaling rule:

E (|X(t)|q ) = c(q)tτ (q)+1 , (3)

where τ (q) and c(q) are both deterministic functions of q.

sure. The resulting process:

good understanding of multifractal measures, which we now present.

3.2 The Binomial Measure is the Simplest Example of a Multifractal

of µ1 is drawn in Figure 4a for m0 = 0.6.

the dyadic set [1/2, 1] and obtain:

µ2 [0, 1/4] = m0 m0 , µ2 [1/4, 1/2] = m0 m1 ,

µ2 [1/2, 3/4] = m1 m0 , µ2 [3/4, 1] = m1 m1 .

Iteration of this procedure generates an infinite sequence of measures. In step k + 1, we assume

where the dyadic number t is of the form:

obtained after k = 4 steps of the recursion.

The measure of the dyadic interval simplifies to:

µ[t, t + 2−k ] = mkϕ0 kϕ1